Professional Documents
Culture Documents
06
Prociclicality of Systemic Credit Risk: The monetary authorities of several countries pro-
A Point-in-Time model for the Brazilian pose a new prudential provision targeted at the sys-
Banking Systems Credit Portfolio Risk temic credit risk. The provision would be created
during economic recovery and peak periods. The
Ana Paula Mussi Szabo Cherobim study assumes the prociclicality of credit risk, that
Jorge Henrique de Frias Barbosa is, the development of credit risk is directly associ-
Wesley Vieira da Silva ated with the macroeconomic environment.
30
Determinants of the Demand for Overdraft The main purpose of the study was to identify the
Limit Credit in the Brazilian Economy factors that determine the demand for personal
credit. To this end, static and dynamic discrete re-
Carlos Eduardo Balbi da Silveira sponse regression models for panel data were used
to model the data of 56,898 overdraft customers of a
Brazilian bank, using balanced information.
54
Directions for Risk Parameter Modeling The article covers the requirements with the great-
within a Basel II and III Context est impact on building testing and validating the in-
ference models used by Internal Risk Rating Sys-
Carlos Antonio Campos Nogueira tems IRRSs, with a focus on the retail and whole-
sale categories, the main targets of the modeling ef-
forts of Financial Institutions.
66
Directions for
Risk Parameter
Modeling Within
a Basel II and III
Context
Carlos
Antonio
Campos
Nogueira
55
Abstract
The Basel accord and its continuity documents
(Basel III) set many rules for the development, validation
and maintenance of Internal Credit Risk Rating Systems,
which are required for financial institutions to apply to IRB
(Internal Ratings Based) under either the basic (FIRB) or
advanced (AIRB) approaches. But the standards leave
gaps, in particular as regards the construction of internal
risk rating systems (IRRS) and the theoretical grounds of
these systems inference engines, which we call models.
The omission is purposeful and intended to give FIs lee-
way to build models and IRRS to best fit their characteris-
tics (customer base, local law, product types, etc.). Within
the range of these rules, FIs have been evolving along sev-
eral lines of operation that this paper will quickly address
within the context of credit risk.
plies the use of means (e.g.: the av- The parameter-mapping rule
erage value of the PDs of the various for pools involves subtleties even in the
obligors in a certain homogeneous segregation of the pools exposures,
risk group, or pool). This method is an which require a delicate compromise
integral part of the IRRS methodolo- between the stability of the ratings and
gy, as are the mathematical models their granularity (number of pools). If
used in inference models. the granularity is too high (small range
IRSS
Qualitative
criteria
Rating A
Rating B
Inference Model
models predictions
Rating H
Other quantitative
criteria
57
Model output
Scores/ratings scale Average parameters parameters
For the models to show good sub-category at the lower sales limit
predictive capacity, it is crucial to cat- (over 3.6M and under 48.6M annually)
egorize the exposures whose data are for collective treatment according to
fed into the model, pooling them ac- pools, as long as the same criterion is
cording to similar risk profiles As a used for the purposes of management
consequence, information from the and pricing.
credit contracts that make up the ex- 2. Quality of the data used
posures must be categorized as stip- for model input variables:
ulated by the Central Bank(Retail/ The variables must be fed
Wholesale). with reliable data from within the
The Retail category must be base for every obligor event in the rel-
divided into the sub-categories Home evant pool. By reliable we mean, for
equity/ Qualified retail revolving cred- example, that static variable data on
it/Other retail exposures. the obligor and/or exposure must vary
Within each retail sub-catego- little or according to trends, and that
ry, IFs may treat exposures according both static and dynamic ones must
to homogeneous risk groups (pools), undergo semantic consistency tests
where ever exposure in a certain pool (e.g.: income does not fluctuate sig-
must have the same values for the risk nificantly for civil servants). This reli-
parameters PD, LGD and EAD (the ability must extend, within the base to
average values associated with each all obligor/exposure events addressed
pool, according to the mapping pro- by the models.
cess seen in Fig. 2). In addition, the domain varia-
Different models exist for each tion of continuous and discrete vari-
risk parameter, that is, there are spe- ables must be small over time. The
cific models for PD, others for LGD absence of syntactic and semantic
and a third type intended to predict variation within the domain must be
EAD. A single PD model may be used proven. To illustrate: a variable that
for one or more pools. In addition, one proposes to locate a retail obligor ac-
or more pools may show the same set cording to annual income range may
of risk parameters as predicted by the do so on a discrete, 6-range domain
respective models. (0-5). That is, not only do the ranges
The Wholesale category must remain stable (they persist syntacti-
be divided into the sub-categories Ex- cally from 0 to 5) but they also retain
posures to individuals not covered by the same meaning(range 3, for exam-
the retail category and small and me- ple, comprehends customers with in-
dium enterprises (SMEs)/Specialized comes in the 30,001-50,000 range).
Financings (Project Finance or Spe- Finally, the independent vari-
cific asset financing)/Other wholesale ables (mode inputs) must prove their
exposures. The FI may, at its discre- explanatory capacity for the obligors
tion, create a sub-division of the SME risk profile.
59
Art. 63. Estimations of the PD, (which may derive from the teams that
LGD and EAD parameters must meet model PD).
the following requirements: However, the decision to polar-
II - taking into consideration all ize LGD may have disadvantages, in-
of the quantitative and qualitative infor- cluding:
mation available, according to the rele- Although admissible as an ap-
vance criterion; proach, depending on the manner and
scope of utilization by an FI, the su-
4.2. On the Use of Binary pervisor may interpret it as a conven-
LGDs as Model Inputs ient simplification, a case of adapt-
One way to address the chal- ing the methodology to the availabili-
lenge of accurately ascertaining ob- ty of (and preference for using) a (Log-
served LGDs (the models dependent it) tool, rather than the outcome of the
input variable) is to use the so-called bi- search for a model with good predictive
nary LGD. powers. Logit works well for PD mod-
The approach polarizes the val- els because of the actual binary na-
ues of input LGDs (for the dependent ture of the observed variable (in default
variable) as either zero or 100%. The de- or performing). For LGD, the actual ob-
cision to round LGD up or down to one served values lie on a continuum that is
of the extremes is usually left to an au- not even limited to the [0,1] interval (al-
tomated routine (program) that uses a though negative LGDs end up being
random method or comparison with a compressed into LGD=0)
cutoff threshold. One must remember that the
It is a valid resource and with argument that LGDs show bimodal dis-
clear benefits, specially: tribution (the grounds for defending ap-
It reduces the time and effort proximation via a binary model) is not
involved in calculating workouts. To il- valid for every portfolio and may create
lustrate: if a defaulting exposure un- distortions (some portfolios are more
dergoing recovery has reached such unimodally distributed and others show
a point that LGD is already below the a shallower U). Using a percentage
threshold, it becomes available as a recovery threshold as a trigger to close
modeling piece of data, as the polarized workout ascertainment may make the
value would already be zero, regardless distribution even more distant from bi-
of any future recoveries. modal/binary behavior.
It approximates the character- The use of binary LGDs as an
istically bimodal distribution of LGD that input may create in wholesale portfoli-
is frequent in some portfolios. os, as a result of leaps in the resulting
It enables using widespread peak distribution (which, in turn, arises
PD modeling tools such as Logit (log from discrete changes due to the LGD
regression), simpifying and expedit- binarization method), large month-
ing the formation of modeling teams to-month allocated capital variations,
65
which may be difficult to explain to su- out-based LGDs into binary values (to
pervisors. There is also the problem of create the dependent variable), using
the compression of negative LGDs thresholds that are consistent with the
into LGD=0 for some portfolios, such respective portfolios/pools.
as real-estate, for example, which tends
to create an LGD frequency distribution 5. Conclusion
that is closer to unimodal, but may con- The diversity and dynamics of
tain topical peaks. topics relative to PD and LGD mode-
For exposures treated as ling have been leading to countless
pools, limiting only two LGD values as a debates, most of which are still on-
dependent variable may restrict diversi- going, in addition to abundant litera-
fication and cause concentrations (as a ture; this article may be just a teas-
result of the leaps described above for er for those with an interest in the
wholesale) that are undesirable to both subject and market practitioners,
the FI and the supervisor. who may do research and gain more
If this approach is selected, cau- in-depth knowledge before making
tion must be taken transforming work- their choices.
Author
Basel Accord International Convergence of Capital Measurement and Capital Standards: A Revised Framework -
Comprehensive Version (June 2006 version): http://www.bis.org/publ/bcbs128.pdf
Central Bank of Brazil Circular Letter No. 3.581, dated mar/08/2012, on the rules for implementing internal credit-risk
models (IRB).
Working Paper 14 Studies on the Validation of Internal Rating Systems: http://www.bis.org/publ/bcbs_wp14.
pdf?noframes=1
Basel II Implementation A guide to developing and validating a compliant, IRRS - Ozdemir, Miu (MC Graw Hill).
Credit Technology Review Serasa Experian No. #74, pp. 31 and 46.
Credit Technology Review Serasa Experian No. #76, p. 71.
Comparison of regression models for LGD estimation
Comparison of regression models for LGD estimation A. Arsova, M. Haralampieva, T. Tsvetanova Experian Decision
analytics 2011.